Abstract: A load balancing method classifies requests routed by a client computer to a server computer as either idempotent or non-idempotent requests. Idempotent requests can be submitted multiple times if not processed on the first attempt. Non-idempotent requests are submitted only once. The method designates a last server that provided an indication that it last processed a request and marks that server as the last available server. In one implementation, the method routes non-idempotent requests to the last available server and routes idempotent requests to any server or any of a subset of servers.